Agent Actions

Actions are the discrete operations executed by an AI agent to interact with its environment. In the context of an llm-wiki-pattern, actions are the mechanism by which the agent modifies the vault or queries external data.

The Stop and Parse Approach

For an agent to function reliably, it must follow a strict lifecycle:

1. Generation: The agent outputs a structured command (JSON or Code).
2. Halting: The LLM stops token generation immediately after the command.
3. Parsing: An external system (the "Env") parses the command, executes the tool, and returns the Observation.

Classification of Actions

• Information Gathering: Web searches, database queries.
• Tool Usage: API calls, calculations.
• Environment Interaction: File system operations (e.g., write_file in this wiki).
• Communication: Inter-agent or user-facing chat.

See Also

• code-agents
• wiki-pattern-operations
• yanp-for-agentic-workflows
• agent-thought-cycle